README.md

metadata

pretty_name: LongMemEval ReMe Cleaned-S
language:
  - en
tags:
  - long-term-memory
  - benchmark
  - question-answering
configs:
  - config_name: default
    data_files:
      - split: test
        path: longmemeval_s_reme_cleaned.json

LongMemEval ReMe Cleaned-S

longmemeval_s_reme_cleaned.json is a corrected version of the LongMemEval Cleaned-S dataset. It keeps the original questions and haystack sessions while replacing the answer and supporting-session ground truth with the reviewed values from final_groundtruth_cleaned_s.json.

The corrections address inaccurate answers and evidence sessions, including cases where evidence occurred after the question time and therefore leaked future information into the evaluation.

Sources

• Upstream dataset: xiaowu0162/longmemeval-cleaned
• Upstream file: longmemeval_s_cleaned.json
• Upstream revision: 98d7416c24c778c2fee6e6f3006e7a073259d48f
• Upstream file SHA-256: d6f21ea9d60a0d56f34a05b609c79c88a451d2ae03597821ea3d5a9678c3a442
• Reviewed ground-truth SHA-256: 97dd2b0df872f12c682041c44fe8082f9fa55ab882f5c78c9237b7c77ba9bb00
• Generated dataset SHA-256: f8b308cc2d184d029dd2226e0c0a88eeac298e6318706d588e015beb1b83fc7d

Transformation

Records are matched one-to-one by question_id. The generated file preserves the upstream schema and applies these mappings:

Output field	Value
answer	Reviewed answer
original_answer	Upstream answer
answer_session_ids	Reviewed evidence_session_ids
original_answer_session_ids	Upstream answer_session_ids

The reviewed evidence_session_ids are written to answer_session_ids so the result remains compatible with the upstream Cleaned-S format. All other fields, including question, question_date, question_type, haystack_sessions, haystack_session_ids, and haystack_dates, are retained from the upstream records.

Change statistics

Changes are counted using exact JSON value equality.

Metric	Records	Share
Total	500	100.0%
Any corrected field changed	77	15.4%
No corrected field changed	423	84.6%
answer changed	57	11.4%
answer_session_ids changed	58	11.6%
Both fields changed	38	7.6%
Only answer changed	19	3.8%
Only answer_session_ids changed	20	4.0%

Evidence-session totals changed from 948 to 819. The corrections removed 129 evidence references across 58 records and added none. Records with an empty evidence list increased from 0 to 58; all 58 corrected records with empty evidence use an answer beginning with The information provided is not enough.

Question type	Total	Answer changed	Evidence changed	Either changed
knowledge-update	78	2	6	8
multi-session	133	7	11	17
single-session-assistant	56	0	0	0
single-session-preference	30	4	0	4
single-session-user	70	1	0	1
temporal-reasoning	133	43	41	47

Validation

• Both inputs contain 500 unique question_id values, with identical ID sets.
• question_type and question time match for all 500 records.
• Every corrected evidence ID exists in the corresponding haystack_session_ids list.
• The generated file contains 500 unique records and retains every non-target value from the upstream dataset.

Citation

If you use this dataset, please cite the original LongMemEval authors:

Wu D, Wang H, Yu W, et al. Longmemeval: Benchmarking chat assistants on long-term interactive memory. arXiv preprint arXiv:2410.10813, 2024.